Page 78 -

52 Chapter 2 ■ Edge-Detection Techniques

suppresses false zero crossings, performs adaptive gradient thresholding, and
finally also applies hysteresis thresholding. In both methods, as with Marr and
Hildreth, the authors suggest the use of multiple resolutions.
Both algorithms offer user-specified parameters, which can be useful for
tuning the method to a particular class of images. The parameters are:

Canny Shen-Castan (ISEF)
Sigma (standard deviation) 0<=b<=1.0 (smoothing factor)
High hysteresis threshold High hysteresis threshold
Low hysteresis threshold Low hysteresis threshold

Width of window for adaptive gradient
Thinning factor

The algorithms were implemented according to the specification laid out in
the original articles describing them. It should be pointed out that the various
parts of the algorithms could be applied to both methods; for example, a thin-
ning factor could be added to Canny’s algorithm, or it could be implemented
using recursive filters. Exploring all possible permutations and combinations
would be a massive undertaking.
Figure 2.16 shows the result of applying the Canny and the Shen-Castan
edge detectors to the test images. Because the Canny implementation uses a
wrap-around scheme when performing the convolution, the areas near the
boundary of the image are occupied with black pixels, although sometimes
with what appears to be noise. The ISEF implementation uses recursive
filters, and the wrap-around was more difficult to implement; it was not, in
fact, implemented. Instead, the image was embedded in a larger one before
processing. As a result, the boundary of these images is mostly white where
the convolution mask exceeded the image.
The two methods were evaluated using E1 and E2, even though flaws have
been found with E1. ISEF seems to have the advantage as noise becomes
greater, at least for the E1 metric, as shown in Table 2.6.
Canny has the advantage using the E2 metric, as shown in Table 2.7.
Overall, the ISEF edge detector is ranked first by a slight margin over Canny,
which is second. Marr-Hildreth is third, followed by Kirsch, Sobel, 2 and 1
in that order. The comparison between Canny and ISEF does depend on the
parameters selected in each case, and it is likely that better evaluations can be
found that use a better choice of parameters. In some of these the Canny edge
detector will come out ahead, and in some the ISEF method will win. The best
set of parameters for a particular image is not known, and so ultimately the
user is left to judge the methods.